Re: [PATCH] ACPICA: fix UBSAN warning using __builtin_offsetof

[Jung-uk Kim]

On 20. 6. 10., Nick Desaulniers wrote:
> On Wed, Jun 10, 2020 at 4:07 PM Kaneda, Erik <erik.kaneda@xxxxxxxxx> wrote:
>>
>> +JKim (for FreeBSD's perspective)
>>
>>> -----Original Message-----
>>> From: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
>>> Sent: Monday, June 1, 2020 4:18 PM
>>> To: Moore, Robert <robert.moore@xxxxxxxxx>; Kaneda, Erik
>>> <erik.kaneda@xxxxxxxxx>; Wysocki, Rafael J <rafael.j.wysocki@xxxxxxxxx>;
>>> Len Brown <lenb@xxxxxxxxxx>
>>> Cc: Ard Biesheuvel <ardb@xxxxxxxxxx>; dvyukov@xxxxxxxxxx;
>>> glider@xxxxxxxxxx; guohanjun@xxxxxxxxxx; linux-arm-
>>> kernel@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
>>> lorenzo.pieralisi@xxxxxxx; mark.rutland@xxxxxxx;
>>> ndesaulniers@xxxxxxxxxx; pcc@xxxxxxxxxx; rjw@xxxxxxxxxxxxx;
>>> will@xxxxxxxxxx; stable@xxxxxxxxxxxxxxx; linux-acpi@xxxxxxxxxxxxxxx;
>>> devel@xxxxxxxxxx
>>> Subject: [PATCH] ACPICA: fix UBSAN warning using __builtin_offsetof
>>>
>>> Will reported UBSAN warnings:
>>> UBSAN: null-ptr-deref in drivers/acpi/acpica/tbfadt.c:459:37
>>> UBSAN: null-ptr-deref in arch/arm64/kernel/smp.c:596:6
>>>
>> Hi,
>>
>>> Looks like the emulated offsetof macro ACPI_OFFSET is causing these. We
>>> can avoid this by using the compiler builtin, __builtin_offsetof.
>>>
>> This doesn't really fly because __builtin_offsetof is a compiler extension.
>>
>> It looks like a lot of stddef.h files do this:
>>
>> #define offsetof(a,b) __builtin_offset(a,b)
>>
>> So does anyone have objections to ACPI_OFFSET being defined to offsetof()?
>>
>> This will allow a host OS project project to use their own definitions of offsetof in place of ACPICA's.
>> If they don't have a definition for offsetof, we can supply the old one as a fallback.
>>
>> Here's a patch:
>>
>> --- a/include/acpi/actypes.h
>> +++ b/include/acpi/actypes.h
>> @@ -504,11 +504,17 @@ typedef u64 acpi_integer;
>>  #define ACPI_SUB_PTR(t, a, b)           ACPI_CAST_PTR (t, (ACPI_CAST_PTR (u8, (a)) - (acpi_size)(b)))
>>  #define ACPI_PTR_DIFF(a, b)             ((acpi_size) (ACPI_CAST_PTR (u8, (a)) - ACPI_CAST_PTR (u8, (b))))
>>
>> +/* Use an existing definiton for offsetof */
> 
> s/definiton/definition/
> 
>> +
>> +#ifndef offsetof
>> +#define offsetof(d,f)                   ACPI_PTR_DIFF (&(((d *) 0)->f), (void *) 0)
>> +#endif
> 
> If this header doesn't explicitly include <stddef.h> or
> <linux/stddef.h>, won't translation units that include
> <acpi/actypes.h> get different definitions of ACPI_OFFSET based on
> whether they explicitly or transitively included <stddef.h> before
> including <acpi/actypes.h>?  Theoretically, a translation unit in the
> kernel could include actypes.h, have no includes of linux/stddef.h,
> then get UBSAN errors again from using this definition?
> 
> I don't mind using offsetof in place of the builtin (since it
> typically will be implemented in terms of the builtin, or is at least
> for the case specific to the Linux kernel). But if it's used, we
> should include the header that defines it properly, and we should not
> use the host's <stddef.h> IMO.  Is there a platform specific way to
> include the platform's stddef.h here?
> 
> Maybe linux/stddef.h should be included in
> include/acpi/platform/aclinux.h, then include/acpi/platform/acenv.h
> included in include/acpi/actypes.h, such that ACPI_OFFSET is defined
> in terms of offsetof defined from that transitive dependency of
> headers? (or do we get a circular inclusion trying to do that?)

Actually, I think we should let platform-specific acfoo.h decide what to
do here, i.e.,

#ifndef ACPI_OFFSET
#define ACPI_OFFSET(d, f) ...
#endif

Jung-uk Kim

>> +
>>  /* Pointer/Integer type conversions */
>>
>>  #define ACPI_TO_POINTER(i)              ACPI_CAST_PTR (void, (acpi_size) (i))
>>  #define ACPI_TO_INTEGER(p)              ACPI_PTR_DIFF (p, (void *) 0)
>> -#define ACPI_OFFSET(d, f)               ACPI_PTR_DIFF (&(((d *) 0)->f), (void *) 0)
>> +#define ACPI_OFFSET(d, f)               offsetof (d,f)
>>  #define ACPI_PHYSADDR_TO_PTR(i)         ACPI_TO_POINTER(i)
>>  #define ACPI_PTR_TO_PHYSADDR(i)         ACPI_TO_INTEGER(i)
>>
>> Thanks,
>> Erik
>>
>>> The non-kernel runtime of UBSAN would print:
>>> runtime error: member access within null pointer of type for this macro.
>>>
>>> Link: https://lore.kernel.org/lkml/20200521100952.GA5360@willie-the-truck/
>>> Cc: stable@xxxxxxxxxxxxxxx
>>> Reported-by: Will Deacon <will@xxxxxxxxxx>
>>> Suggested-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
>>> Signed-off-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
>>> ---
>>>  include/acpi/actypes.h | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/include/acpi/actypes.h b/include/acpi/actypes.h index
>>> 4defed58ea33..04359c70b198 100644
>>> --- a/include/acpi/actypes.h
>>> +++ b/include/acpi/actypes.h
>>> @@ -508,7 +508,7 @@ typedef u64 acpi_integer;
>>>
>>>  #define ACPI_TO_POINTER(i)              ACPI_CAST_PTR (void, (acpi_size) (i))
>>>  #define ACPI_TO_INTEGER(p)              ACPI_PTR_DIFF (p, (void *) 0)
>>> -#define ACPI_OFFSET(d, f)               ACPI_PTR_DIFF (&(((d *) 0)->f), (void *)
>>> 0)
>>> +#define ACPI_OFFSET(d, f)               __builtin_offsetof(d, f)
>>>  #define ACPI_PHYSADDR_TO_PTR(i)         ACPI_TO_POINTER(i)
>>>  #define ACPI_PTR_TO_PHYSADDR(i)         ACPI_TO_INTEGER(i)
>>>
>>> --
>>> 2.27.0.rc2.251.g90737beb825-goog